A Protocol to Evaluate Carbon Dioxide Flux Partitioning Methods for Eddy Covariance Data

Carbon dioxide (CO₂) flux partitioning involves separating net ecosystem exchange (NEE) into its gross primary production (GPP) and ecosystem respiration (RECO) components. Despite 25 years of flux research and abundant data from networks such as FLUXNET [1], the development and validation of new partitioning methods remain hindered by the lack of a standardized benchmark dataset and evaluation protocol. Existing parametric methods, including nighttime (NT) [2] and daytime (DT) [3] approaches, have become integral to FLUXNET data products but face limitations such as dependency on assumptions and variable robustness across biomes and conditions. Emerging machine-learning (ML)-based methods offer flexibility and reduced reliance on assumptions but require rigorous evaluation [4,5,6].

We establish a benchmark dataset and standardized evaluation protocol to address these challenges. The dataset includes synthetic data generated by multiple mechanistic models, allowing comparison against a known ground truth. These models simulate diverse biomes and environmental conditions, including rapid system changes and extreme events. Additionally, the dataset incorporates realistic data gaps and noise scenarios to test method resilience. The evaluation includes multiple performance metrics across different temporal scales. We assess the ability of methods to capture critical meteorological events and ecological transitions.

Our results indicate that for GPP, ML methods outperform parametric methods at half-hourly scales and in capturing daily anomalies, though the extent of improvement depends on the setup of the ML method. Conversely, NT method performs better at representing the monthly diurnal cycle and seasonal trends. For RECO, the NT method yields the most robust overall results but struggles to capture sudden changes in ecosystem dynamics, which ML methods handle more effectively. Across all methods, daily anomalies remain a persistent challenge, highlighting the need for dynamic ML models. Furthermore, we find that NEE data availability below approximately 30% for a site-year reduces the reliability of the current neural network methods, suggesting the need for transfer or meta-learning schemes or improved gap-filling strategies.

This initiative streamlines the development and comparison of partitioning methods, enabling transparent assessment of their strengths and weaknesses.

References: 

[1] Baldocchi, Dennis, et al. “FLUXNET: A new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities.” Bulletin of the American Meteorological Society 82.11 (2001): 2415-2434.

[2] Reichstein, Markus, et al. "On the separation of net ecosystem exchange into assimilation and ecosystem respiration: review and improved algorithm. Global Change Biology 11.9 (2005): 1424-1439. https://doi.org/10.1111/j.1365-2486.2005.001002.x

[3] Lasslop, Gitta, et al. “Separation of net ecosystem exchange into assimilation and respiration using a light response curve approach: critical issues and global evaluation.” Global Change Biology 16.1 (2010): 187–208. https://doi.org/10.1111/j.1365-2486.2009.02041.x

[4] Tramontana, Gianluca, et al. “Partitioning net carbon dioxide fluxes into photosynthesis and respiration using neural networks.” Global change biology 26.9 (2020): 5235-5253. https://doi.org/10.1111/gcb.

[5] Zhan, Weiwei, et al. “Two for one: Partitioning co2 fluxes and understanding the relationship between solar-induced chlorophyll fluorescence and gross primaryp roductivity using machine learning.” Agricultural and Forest Meteorology 321 (2022): 108980. https://doi.org/10.1016/j.agrformet.2022.108980

[6] Kai-Hendrik Cohrs et al. “Causal hybrid modeling with double machine learning—applications in carbon flux modeling.” Machine Learning: Science and Technology 5 (2024): 035021. https://doi.org/10.1088/2632-2153/ad5a60